search

toshikwa / gail-airl-ppo.pytorch

PyTorch implementation of GAIL and AIRL based on PPO.
MIT License
189 stars 30 forks source link

About the parameter rollout_length settings #9

Open nicholas0717 opened 1 year ago

nicholas0717 commented 1 year ago

Hi @toshikwa :

I'm trying to implement other tasks with your GAIL framework, but it works badly. I think the network units and relevant parameters I chose are not appropriate. I have some questions.

I want to know how to choose the rollout_lengthfor different tasks. You chose 2000 for InvertedPendulum-v2 and 50000 for Hopper-v3. Is there a criterion for choosing rollout_length?

The hidden_activation you chose are all Tanh() in discriminator, actor, critic. Is it better than ReLU() in GAIL? And will the hidden_units affect the final performance of the agent a lot?

Last question. In your Hopper-v3 example, the finalacc_exp and acc_pi is shown below. But in my task, the Accuracy Exp is approaching 1. Is there something wrong with this？

image image

Hope you can give me some advice. Thank you so much.

Charlesyyun commented 1 year ago

Hi nicholas, I have also tried the code on my task. It turns out that the results are unsatisfying. Did you find a way out?

JimberZ commented 1 year ago

Hi nicholas, I have also tried the code on my task. It turns out that the results are unsatisfying. Did you find a way out?

so do i, Did you find a way out?